基于迁移学习的电力系统暂态稳定自适应预测  

Self-adaptive Power System Transient Stability Prediction Based on Transfer Learning

作　　者：张若愚 吴俊勇[1] 李宝琴 邵美阳 ZHANG Ruoyu;WU Junyong;LI Baoqin;SHAO Meiyang(School of Electrical Engineering,Beijing Jiaotong University,Haidian District,Beijing 100044,China)

机构地区：[1]北京交通大学电气工程学院,北京市海淀区100044

出　　处：《电网技术》2020年第6期2196-2203,共8页Power System Technology

基　　金：国家重点研发计划项目(2018YFB0904500);国家电网有限公司科技项目(SGLNDK00KJJS1800236)。

摘　　要：基于人工智能的电力系统暂态稳定预测,通常需要用离线生成的大量暂稳样本对预测模型进行训练,然后根据系统的实时响应进行在线预测。但当系统的运行方式和拓扑结构发生较大变化时,预测模型的精度会显著下降,亟需一种能跟踪系统变化的自适应暂稳预测方法。针对该问题,将迁移学习引入电力系统暂稳预测,基于卷积神经网络提出了一种自适应预测方法。首先利用离线生成的大量暂稳样本训练并得到基于卷积神经网络的预训练模型。当系统运行方式和拓扑结构发生较大变化时,保持预训练模型的网络结构不变,将其中的2个卷积层、2个池化层和全连接层的网络参数迁移至新模型;提出了一种最小均衡样本集的变步长生成方法,用新生成的最小均衡样本集训练分类层参数,从而快速得到新的预测模型。新英格兰10机39节点系统的测试结果表明:所提方法能自适应跟踪系统运行方式和拓扑结构的变化,有效更新预测模型且大幅减少新模型的训练时间,为基于人工智能的电力系统暂态稳定自适应预测提供了一条新思路。Transient stability prediction of power systems based on artificial intelligence usually requires training a prediction model with a large number of samples generated offline,and then performing online prediction based on the real-time response of the system.However,when the system’s operating mode and topological structure change greatly,the accuracy of the prediction model will decrease significantly.Therefore,a self-adaptive transient stability prediction method that can track the changes of the system is urgently needed.Considering this problem,transfer learning is introduced into transient stability prediction,and a self-adaptive prediction method is proposed based on convolutional neural networks.Firstly,a pre-trained model is obtained based on the convolutional neural network by training a large number of samples generated at the offline stage.When the operating mode and the topological structure change greatly,the network structure of the pre-trained model can be kept unchanged and the network parameters in the two convolutional layers,the two pooling layers,and the fully connected layer are transferred into the new model.A variable step dataset generation method is used to retrain the parameters of classification layer with a minimum balanced dataset so that an updated model can be achieved quickly.Experiment results in the New England 10-machine 39-bus system demonstrate that the proposed method can effectively update the prediction model and reduce training time dramatically.It provides a new idea for selfadaptive prediction of power system transient stability based on artificial intelligence.

关 键 词：迁移学习 深度学习 卷积神经网络 电力系统 暂态稳定预测 

分 类 号：TM721[电气工程—电力系统及自动化]